Method and system for adaptively searching alcoholization degree of inventory flue-cured tobacco lamina

Abstract

The invention relates to the technical field of tobaccos, and provides a method and a system for adaptively searching the alcoholization degree of inventory lamina. The method comprises the following steps: acquiring near infrared spectrum data and corresponding alcoholization degree of stock tobacco lamina, and performing feature extraction on the near infrared spectrum data of the stock tobacco lamina to obtain a plurality of spectrum feature variables; taking the plurality of spectral characteristic variables as input of a flue-cured tobacco lamina alcoholization degree discrimination model, taking the corresponding alcoholization degree as output, training the flue-cured tobacco lamina alcoholization degree discrimination model, and calculating a target function value according to a model output result and the corresponding alcoholization degree; a spectral variable selection algorithm is adopted, the spectral characteristic variables meeting the requirements are screened according to the objective function value and the evaluation of the advantages and disadvantages of the spectral characteristic variables, and the lamina alcoholization degree discrimination model is input again for iterative training until the set requirements are met; and based on the near infrared spectrum data of the to-be-detected tobacco lamina, obtaining the alcoholization degree of the tobacco lamina by adopting a spectrum variable selection algorithm and the tobacco lamina alcoholization degree judgment model.

Description

Technical Field

[0001] The present invention relates to the technical field of tobacco, and in particular to a method and system for adaptively searching for alcoholization degree determination of stock tobacco strips. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Tobacco leaf raw material is the most important material foundation for the quality of cigarette products. Raw tobacco leaves undergo redrying and become tobacco sheets, the primary raw material for cigarette production. After field growth, harvesting, curing, and redrying, tobacco leaves commonly exhibit issues such as inconsistent internal chemical properties, a strong green odor, and a less mellow flavor. Their quality does not meet the requirements for high-quality cigarette production and cannot be directly used in cigarette production. To improve tobacco leaf usability, the industry often ages the redrying leaves to reduce inherent quality defects, meet industrial production needs, and maximize the utilization of the raw material. Tobacco companies use natural aging to improve and enhance the quality of redrying tobacco leaves. The raw tobacco leaves typically undergo two to three years of natural aging in warehouses before they can be used in production. Due to the long aging cycle, some leaves darken in color and become over-aged during prolonged storage, resulting in a decline in quality and affecting their usability. During the tobacco leaf aging process, tobacco companies collect samples from the aging warehouse and conduct sensory evaluations. Only when the sensory quality of the samples meets the scoring standards can they be used in production. However, the traditional sensory evaluation method first prepares the aged tobacco leaves through a complex process of cutting, drying, and rolling into cigarette samples. Then, no fewer than seven sensory evaluation experts are invited to evaluate and score them according to sensory quality inspection standards. On the one hand, sensory evaluation experts are required to acquire extensive sensory quality evaluation technical experience through repeated training and possess an excellent memory to solidify their perception of product quality characteristics. This requires a very high level of technical expertise and a long training cycle. On the other hand, relying on repeated evaluations by experts makes the evaluation process time-consuming and costly. Furthermore, due to the influence of many factors such as the experts' personal preferences, technical level, and physical condition, the quality perception of the same product can sometimes vary significantly between different experts, and even within the same expert at different times, making product quality evaluation highly subjective. Therefore, the traditional method of determining the degree of aging of aged tobacco strips has problems such as long cycle, high cost, and low efficiency. It is also affected by the status of experts and personal preferences. The smoking evaluation results are highly subjective and the evaluation is unstable, resulting in a lack of credibility in the quality determination results of aged tobacco strips. Summary of the Invention

[0004] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a method and system for adaptively searching for the alcoholization degree of inventory tobacco slices. The method of the present invention can obtain accurate and objective classification results through near-infrared spectroscopy detection and model judgment method.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The first aspect of the present invention provides a method for adaptively searching for the degree of aging of tobacco chips in stock.

[0007] A method for adaptively searching for alcohol-aging degree of tobacco chips in stock, comprising:

[0008] Acquire near-infrared spectral data and corresponding degrees of aging of tobacco strips in stock, perform feature extraction on the near-infrared spectral data of the tobacco strips in stock, and obtain a number of spectral characteristic variables; use the several spectral characteristic variables as inputs to a tobacco strip aging degree discrimination model, with the corresponding degrees of aging as outputs, train the tobacco strip aging degree discrimination model, and calculate the objective function value based on the model output results and the corresponding degrees of aging; use a spectral variable selection algorithm to evaluate the quality of the spectral characteristic variables based on the objective function value, select spectral characteristic variables that meet the requirements, and re-input them into the tobacco strip aging degree discrimination model for iterative training until the objective function value meets the set requirements;

[0009] Feature extraction is performed on the acquired near-infrared spectral data of the tobacco strips to be tested to obtain a number of spectral characteristic variables to be tested; a spectral variable selection algorithm is used to screen the number of spectral characteristic variables to be tested, and the screened spectral characteristic variables to be tested are input into a tobacco strip aging degree discrimination model to obtain the aging degree of the tobacco strips.

[0010] Furthermore, the process of constructing the model for distinguishing the degree of aging of tobacco strips includes: setting a search space for a cellular structure, selecting the cellular structure with the best performance and training it until convergence, obtaining a set of search parameters for the optimal structure, recording the optimal weights, and assigning them to all combined cellular structures in a one-dimensional convolutional neural network. Using a cellular structure, hierarchically searching and expanding layer by layer in the superimposed and combined search spaces to construct a model for distinguishing the degree of aging of tobacco strips.

[0011] Furthermore, the search space of the cellular structure is set, the cellular structure with the best performance is selected and trained until convergence, a search parameter set of the optimal structure is obtained, the optimal weight is recorded, and the optimal weight is assigned to all combinations of cellular structures in the one-dimensional convolutional neural network; the method includes:

[0012] The search space of the cellular structure is formed by the value ranges of three hyperparameters: the convolution kernel size n1, the size of the downsampling kernel, and the number of feature maps N;

[0013] Normalize the value range of each hyperparameter to [0,1], randomly select a set of parameter configurations, and calculate the objective function EI(θ);

[0014] In order to deal with the mixed parameter space search range, a local neighborhood search range is initialized around the set of parameters, and the parameter values ​​θ′ of the S nearest neighbors are selected from the univariate Gaussian distribution as neighborhood samples;

[0015] For the S nearest neighbor samples, evaluate their objective function EI(θ′) at one time and select the neighbor sample parameter θ corresponding to the optimal objective function value inc , continue searching with it as the center;

[0016] Repeat the local search process and select the maximum EI(θ inc ) corresponds to the cell structure parameter θ inc , record the optimal cell structure weight ω;

[0017] According to the obtained optimal hyperparameter combination, the cellular structure is expanded into a one-dimensional convolutional neural network.

[0018] Furthermore, the method uses a cellular structure to hierarchically perform layer-by-layer search expansion in the superimposed and combined search space to construct a model for distinguishing the degree of aging of tobacco leaves; the method includes:

[0019] Select the optimal cell structure weight ω obtained by the search, add it to the queue, and share this weight in the next layer;

[0020] Starting from the second layer, the model of each layer is expanded by adding all possible cell structures in the allowed search space, and all the models expanded in this layer are trained and evaluated;

[0021] The two network structures with the best prediction performance are selected and expanded into a complete one-dimensional convolutional neural network, which is added to the next layer queue; the network structure is retrained using the training data set and the weights of the previous layers are updated;

[0022] The cellular structure is added layer by layer from the second layer to the tenth layer, and the above steps are repeated until the preset search termination condition is reached to obtain a model for distinguishing the degree of alcoholization of tobacco leaves.

[0023] Furthermore, the spectral variable selection algorithm is used to evaluate the quality of the spectral characteristic variables according to the objective function value, and to screen the spectral characteristic variables that meet the requirements; the method includes:

[0024] Initialize the harmony memory HM, initialize each component in the harmony vector, and record the objective function value of each component to form a harmony memory HMS×(n+1);

[0025] Compare the random number r1 between [0,1] and the probability HMCR of the harmony memory library to decide whether to select the harmony vector from the harmony memory library; compare the random number r2 between [0,1] and the pitch fine-tuning probability PAR, adjust the harmony vector according to the pitch fine-tuning bandwidth BW adaptive adjustment method, and generate a new harmony vector X new ;

[0026] According to the update of the Sigmoid function after fine-tuning the harmony variable, a random number r4 is generated between [0-1] and Sigmoid(x′ i ) for comparison, if r4<Sigmoid(x′ i ), x′ i The value is set to 1; if r4≥Sigmoid(x′ i ), x′ i The value of is set to 0 to represent each dimension of the harmony vector only with 0 or 1;

[0027] Evaluate the newly generated harmony vector X new , if f(X new ) is better than f(X worst ), then change the X in HM worst With the new harmony vector X new Replace; otherwise, X new Discard it directly, keep HM unchanged, and repeat the iteration until the number of iterations reaches Tmax.

[0028] Furthermore, the objective function value is calculated based on the model output result and the corresponding degree of aging; the method includes: calculating the mean classification accuracy rate as the objective function value based on the model output result and the corresponding degree of aging.

[0029] A second aspect of the present invention provides a system for adaptively searching and determining the degree of alcoholization of tobacco chips in stock.

[0030] A self-adaptive search and aging degree determination system for tobacco chips in stock, comprising:

[0031] The model training module is configured to: obtain near-infrared spectral data and corresponding degrees of aging of inventory tobacco strips, perform feature extraction on the near-infrared spectral data of the inventory tobacco strips to obtain a number of spectral characteristic variables; use the several spectral characteristic variables as inputs to a tobacco strip aging degree discrimination model, use the corresponding degrees of aging as outputs, train the tobacco strip aging degree discrimination model, and calculate an objective function value based on the model output results and the corresponding degrees of aging; use a spectral variable selection algorithm to evaluate the quality of the spectral characteristic variables based on the objective function value, screen the spectral characteristic variables that meet the requirements, and re-input the spectral characteristic variables into the tobacco strip aging degree discrimination model for iterative training until the objective function value meets the set requirements;

[0032] The model prediction module is configured to: extract features from the acquired near-infrared spectral data of the tobacco strips to be tested to obtain a number of spectral characteristic variables to be tested; use a spectral variable selection algorithm to screen the number of spectral characteristic variables to be tested, input the screened spectral characteristic variables to be tested into the tobacco strip aging degree discrimination model to obtain the tobacco strip aging degree.

[0033] A third aspect of the present invention provides a computer-readable storage medium.

[0034] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for adaptively searching for alcoholization degree of stock tobacco sheets as described in the first aspect above.

[0035] A fourth aspect of the present invention provides a computer device.

[0036] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for determining the degree of alcoholization of tobacco leaves in inventory by adaptively searching the method are implemented as described in the first aspect above.

[0037] A fifth aspect of the present invention provides a computer program product or computer program.

[0038] The present invention provides a computer program product or computer program, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the method for determining the degree of alcoholization of tobacco leaves by adaptively searching inventory, as described in the first aspect above.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The present invention provides a method and system for adaptively searching for the alcoholization degree of tobacco strips in stock, which replaces the traditional alcoholization degree determination method of tobacco strips in stock that relies on smoking expert evaluation, and can provide comprehensive, real-time and objective alcoholization degree of tobacco strips.

[0041] This method uses near-infrared equipment to automatically collect spectral data, eliminating the complex processes of cutting, drying, and rolling into cigarettes. This approach is time-efficient, highly efficient, and low-cost, meeting the quality assessment needs of large-scale production enterprises. Compared to the subjectivity inherent in traditional sensory evaluation methods involving multiple experts, this method is more objective and produces more stable results. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0043] Figure 1 This is a flow chart of the method for adaptively searching for alcoholization degree of tobacco chips in stock according to the present invention;

[0044] Figure 2 It is the original spectrum graph shown in the present invention;

[0045] Figure 3 is a pre-processed spectrum diagram shown in the present invention;

[0046] Figure 4 Schematic diagram of the cell structure and one-dimensional convolutional neural network structure shown in the present invention;

[0047] Figure 5 Schematic diagram of the cellular structure convergence results shown in the present invention; wherein (a) is a schematic diagram of the prediction accuracy; (a) is a schematic diagram of the mean square error;

[0048] Figure 6 Schematic diagram of the one-dimensional convolutional neural network structure layer composition shown in the present invention;

[0049] Figure 7 Schematic diagram of the layer-by-layer structural search process shown in the present invention;

[0050] Figure 8 Schematic diagram of the contribution of near-infrared spectroscopy variables categories shown in the present invention;

[0051] Figure 9 Schematic diagram of the matrix structure of the harmony memory (HM) shown in the present invention;

[0052] Figure 10 Schematic diagram of the objective function solution and feedback process shown in the present invention;

[0053] Figure 11 This is a schematic diagram showing the changing trend of classification accuracy as the variable increases;

[0054] Figure 12 It is a structural diagram of the adaptive search inventory tobacco aging degree discrimination system shown in the present invention. DETAILED DESCRIPTION

[0055] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0056] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0057] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0058] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the methods and systems according to the various embodiments of the present disclosure. It should be noted that each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the module, program segment, or a part of code can include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the flowchart and / or block diagram, and the combination of the boxes in the flowchart and / or block diagram, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or can be implemented using a combination of dedicated hardware and computer instructions.

[0059] Explanation of terms:

[0060] Near-infrared spectroscopy primarily reflects the molecular composition and content of organic compounds through the absorption of the harmonic and summed frequencies of hydrogen-containing group vibrations, providing a wealth of quality information. Tobacco near-infrared spectroscopy offers comprehensive quality information, such as flavor characteristics and sensory quality, and is a more effective means of quality testing and evaluation than traditional evaluation techniques.

[0061] Example 1

[0062] like Figure 1As shown, this embodiment provides a method for adaptively searching for the degree of aging of tobacco slices in stock. This embodiment uses the method applied to a server as an example for illustration. It is understandable that the method can also be applied to a terminal, and can also be applied to a system including a terminal, a server, and a server, and is implemented through the interaction between the terminal and the server. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communications, middleware services, domain name services, security services CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected by wired or wireless communication, which is not limited in this application. In this embodiment, the method includes the following steps:

[0063] Acquire near-infrared spectral data and corresponding degrees of aging of tobacco strips in stock, perform feature extraction on the near-infrared spectral data of the tobacco strips in stock, and obtain a number of spectral characteristic variables; use the several spectral characteristic variables as inputs to a tobacco strip aging degree discrimination model, with the corresponding degrees of aging as outputs, train the tobacco strip aging degree discrimination model, and calculate the objective function value based on the model output results and the corresponding degrees of aging; use a spectral variable selection algorithm to evaluate the quality of the spectral characteristic variables based on the objective function value, select spectral characteristic variables that meet the requirements, and re-input them into the tobacco strip aging degree discrimination model for iterative training until the objective function value meets the set requirements;

[0064] Feature extraction is performed on the acquired near-infrared spectral data of the tobacco strips to be tested to obtain a number of spectral characteristic variables to be tested; a spectral variable selection algorithm is used to screen the number of spectral characteristic variables to be tested, and the screened spectral characteristic variables to be tested are input into a tobacco strip aging degree discrimination model to obtain the aging degree of the tobacco strips.

[0065] The present invention selects high-quality spectral characteristic variables through a spectral variable selection algorithm, and improves the accuracy of distinguishing the degree of aging of tobacco strips by inputting the high-quality spectral characteristic variables into a constructed model for distinguishing the degree of aging of tobacco strips.

[0066] In some embodiments, the method comprises obtaining near-infrared spectral data and corresponding aging degrees of the tobacco strips in stock, performing feature extraction on the near-infrared spectral data of the tobacco strips in stock, and obtaining a plurality of spectral characteristic variables; using the plurality of spectral characteristic variables as inputs to a tobacco strip aging degree discrimination model, using the corresponding aging degrees as outputs, training the tobacco strip aging degree discrimination model, and calculating an objective function value based on the model output results and the corresponding aging degrees; using a spectral variable selection algorithm, evaluating the quality of the spectral characteristic variables based on the objective function value, screening the spectral characteristic variables that meet the requirements, and re-inputting the spectral characteristic variables into the tobacco strip aging degree discrimination model for iterative training until the objective function value meets the set requirements; the method comprises:

[0067] Step (1), standard sample collection: 7,725 tobacco leaf samples of different aging states in the warehouse are collected as a standard sample set for determining aging state, and aging degree determination personnel are organized to determine the aging degree of the samples in the standard sample set according to GB / T23220.2-2023.

[0068] Step (2), near-infrared spectrum acquisition and spectrum preprocessing: Use a Vispec QuasIR 2000P near-infrared spectrometer to collect the spectrum data of all the alcoholization state determination standard sample sets in step (1), with a spectrum scanning range of 4000-12800 cm-1 and a resolution of 16 cm -1 , scanning times: 32 times; the first-order derivative + Norris (13) smoothing method was used to preprocess the near-infrared spectrum to eliminate operational errors, background noise, etc., and improve the data signal-to-noise ratio. The original spectrum of the sample and the spectrum after preprocessing are shown in Figure 2 and Figure 3 shown.

[0069] Step (3), training set, validation set, test set and their subset division: All samples of the alcoholization state determination standard sample set were randomly divided according to the allocation principle of 60% for training, 20% for validation, and 20% for testing to form training set, validation set and test set. The overall sample data distribution is shown in Table 1:

[0070] Table 1 Distribution of near infrared spectroscopy sample data

[0071]

[0072]

[0073] Furthermore, when using the joint model for feature variable selection, the training set and validation set are randomly divided into three subsets of roughly equal size to enhance the generalization ability of the variable selection algorithm. The division of the three subset sample data for the joint model is shown in Table 2:

[0074] Table 2. Division of sample data in the joint model subset

[0075]

[0076] Step (4), cellular structure optimization search and verification: Optimize the search for the cellular structure, select the cellular structure model with the best performance and train it until convergence, record its corresponding weights, and assign them to all combined cellular structures in the convolutional neural network. The specific steps are as follows:

[0077] Step (4-1) Cellular structure search space setting: The cellular structure search space is composed of the value ranges of three hyperparameters: the convolution kernel size n1, the downsampling kernel size, and the number of feature maps N.

[0078] The convolution kernel size must be greater than 1. An even-sized convolution kernel cannot guarantee that the input and output sizes of the feature map remain unchanged. The smaller the convolution kernel size, the smaller the amount of calculation and the number of parameters. The range of the convolution kernel size n1 is set to {3, 5, 7, 9, 11, 13, 15, 17}, that is:

[0079] ●1×3 convolution kernel●1×5 convolution kernel●1×7 convolution kernel●1×9 convolution kernel

[0080] ●1×11 convolution kernel●1×13 convolution kernel●1×15 convolution kernel●1×17 convolution kernel

[0081] For each pooling layer, the downsampling kernel n2 range is set to {2, 3, 4, 5}, that is:

[0082] ●1×2 downsampling kernel●1×3 downsampling kernel●1×4 downsampling kernel●1×5 downsampling kernel If the number of feature maps is too small, some features that are beneficial to network learning will be ignored. When the number of feature maps is too large, not only will the training time of the model be greatly increased, but the model will also be prone to overfitting. After comprehensive analysis, the value range of the number of feature maps N is set between [1-20].

[0083] Step (4-2) Cellular structure optimization search: Optimize the search for the cellular structure. First, set the search space F of the cellular structure. On this basis, select the cellular structure model with the best performance and train it until convergence to obtain the search parameter set θ of the optimal structure. inc , record the corresponding weights, and assign them to all combined cell structures in the convolutional neural network. The specific steps are as follows:

[0084] Step (4-2-1): Normalize the value range of each numerical parameter to [0,1], randomly select a set of parameter configurations, and calculate the objective function EI(θ). The calculation formula of the objective function EI(θ) is as follows:

[0085]

[0086] in, represents the cumulative distribution function of the standard normal distribution; f min =μ(θ inc )+σ(θ inc ),θ inc Indicates the optimal parameters selected in the current iteration.

[0087] Step (4-2-2): To process the mixed parameter space search range, initialize a local neighborhood search range around the set of parameters, and select the 4 nearest neighbor parameter values ​​θ′ as neighborhood samples from a univariate Gaussian distribution with mean v and standard deviation 0.2.

[0088] Step (4-2-3): Evaluate the objective function EI(θ′) for the four nearest neighbor samples at once. Select the nearest neighbor sample parameter θ corresponding to the optimal objective function value. inc , and continue searching with it as the center.

[0089] Step (4-2-4): Repeat steps (4-2-2) and (4-2-3). Once the EI(θ′) of no neighboring sample becomes larger (no positive expected improvement), stop each local search and select the maximum EI(θ inc ) corresponds to the cell structure parameter θ inc , record the weight ω of the cell structure.

[0090] Step (4-3) Cellular structure performance evaluation: The optimal cell structure hyperparameter θ obtained by searching in step (4-2) inc , (1) convolution kernel size: 1×11; (2) sampling kernel size: 1×2; (3) number of feature maps: 12, according to its parameter values, the cell structure is expanded into a small convolutional neural network, after expansion, the convolutional neural network is as follows Figure 4 shown.

[0091] The network structure was trained using 4,635 training samples, divided into 103 batches, with 45 samples in each batch; 1,545 validation samples were used for parameter adjustment, and 50 iterations of training were performed, with the learning rate set to 0.01.

[0092] After 30 complete iterations of training on all sample data (corresponding to 3,090 training batches), although the prediction accuracy is not high, the cellular structure model has been able to converge. Therefore, this cellular structure can be used to construct a complete one-dimensional convolutional neural network. The prediction accuracy and prediction mean square error of the cellular structure on the test set are as follows: Figure 5 As shown, (a) represents the prediction accuracy and (b) represents the mean square error.

[0093] Step (5), Construction of a Model for Discriminating the Aging Degree of Tobacco Strips: Based on a one-dimensional convolutional neural network structure adaptive search method that advances layer by layer based on a cellular structure, the simplest cellular structure is first used, and then layer-by-layer search expansion is performed in the superimposed and combined search space to construct a model for distinguishing the aging degree of tobacco strips. The specific steps are as follows:

[0094] Step (5-1): Select the optimal cell structure weight ω obtained in step (4), add it to the queue, and share this weight in the next layer.

[0095] Step (5-2): Starting from the second layer, the model of each layer is expanded by adding all possible cell structures in the allowed search space, and all models expanded in this layer are trained and evaluated.

[0096] Step (5-3): Select the two network structures with the best prediction performance (Top-2) and expand them into a complete one-dimensional convolutional neural network, and add them to the next layer queue; apply the training data set to retrain the network structure and update the weights of the previous layers.

[0097] Step (5-4): Add cellular structures layer by layer from the second layer to the tenth layer, repeat the above steps until the preset search termination condition is reached, and the final one-dimensional convolutional neural network structure is formed. Figure 6 As shown (only the cell structure layer is listed).

[0098] Figure 7 A demonstration process of layer-by-layer search is given.

[0099] Step (6), adaptive selection and verification of near-infrared spectral characteristic variables: the tobacco aging degree discrimination model is combined with the spectral variable selection algorithm, and the target model continuously iteratively evaluates the quality of the selected variables, thereby adaptively selecting the characteristic variables of the near-infrared spectrum and effectively screening the near-infrared spectral characteristic variables, thereby improving the robustness and generalization ability of the model.

[0100] On the training sample set, the KW test method is used to calculate the significance of each spectral variable to the category difference, and then its characteristic contribution is obtained. The interval of 3800cm-1 to 3900cm-1 corresponds to the initial range of spectral acquisition. Since the operation of various optical devices is not stable yet, it often causes large detection errors. When analyzing the spectrum, this interval is usually removed and the contribution of the spectral variables in this interval is set to 0. Figure 8 shown.

[0101] The specific steps for adaptive selection of near-infrared spectral characteristic variables are as follows:

[0102] Step (6-1): The objective function for the spectral feature variable selection is set to the average classification accuracy (AvgAccuracy). The dimension of the near-infrared spectrum is 1609, and the parameters HMS are set to 100, HMCR to 0.95, BW to 0.02, and the number of iterations Tmax to 1,000.

[0103] Step (6-2): Initialize the harmony memory HM. Initialize each component in the harmony vector using the following formula:

[0104]

[0105] Among them, x i represents the i-th component randomly initialized by the original algorithm; g i represents the category contribution index of the i-th component calculated using the KW (Kruskal-Wallis) test method; x′ i Represents the i-th initialization component generated by the improved method. And record its objective function value to form a HMS×(n+1) harmony memory bank. The harmony memory bank (HM) matrix structure is as follows Figure 9 shown.

[0106] Step (6-3): Generate a new harmony vector.

[0107] Step (6-3-1): Compare the random number r1 between [0,1] and the harmony library value probability HMCR to decide whether to select a harmony vector from the harmony memory library. The formula is as follows:

[0108]

[0109] Step (6-3-2): Compare the random number r2 between [0,1] and the pitch fine-tuning probability PAR, and adjust the harmony vector according to the pitch fine-tuning bandwidth BW adaptive adjustment method. The pitch fine-tuning bandwidth BW adaptive adjustment method calculation formula is as follows:

[0110]

[0111] in, and They represent the i-th dimension component of the best and worst harmony variables in the current iteration cycle respectively; λ represents the BW adjustment coefficient, which is dynamically adjusted as the number of iterations changes. The specific expression is as follows:

[0112]

[0113] The formula for adjusting the harmony vector is as follows:

[0114]

[0115] Where r3 is a random number between [0-1].

[0116] Step (6-4): Update the harmony variable. According to the Sigmoid function after fine-tuning the harmony variable, a random number r4 is generated between [0-1] and Sigmoid (x′ i ) for comparison, if r4≤Sigmoid(x′ i ), x′ i The value is set to 1; if r4≥Sigmoid(x′ i ), x′ i The value of is set to 0. Thus, each dimension of the harmony vector is represented by only 0 or 1 (when the feature variable is selected, 1 indicates that the variable is selected, and 0 indicates that the variable is not selected). The calculation process is as follows:

[0117]

[0118] Step (6-5): Objective function solution and feedback. For each divided training subset data, based on the characteristic spectral points selected corresponding to each solution vector, the one-dimensional convolutional neural network constructed in step (5) is used to perform feature extraction and classification prediction, and the mean classification accuracy of each training subset model is calculated as the objective function value. The process is as follows: Figure 10 shown.

[0119] Step (6-6): Evaluate the newly generated harmony vector X new If f(X new ) is better than f(X worst ), then change the X in HM worst With the new harmony vector X new Replace; otherwise, X new Directly discard, HM remains unchanged. Repeat c), d), and e) until the number of iterations reaches Tmax.

[0120] The algorithm was run 50 times on each of the three training subsets, and the cumulative frequency of 1,609 spectral variables being selected in a total of 150 training sessions was counted. The higher the frequency, the more important the spectral variable is to the prediction indicator, and it can be selected as a feature variable. The cumulative frequency of selection of each variable was sorted from high to low, and the number of variables was continuously increased with a descending gradient of 25 frequencies. The classification accuracy on the training set was calculated separately. At the beginning, as the number of variables increased, the accuracy gradually increased, and reached the maximum value when the frequency of variable selection was 100. If the number of variables continues to increase, noise and redundant information will be introduced, and the accuracy will decrease, which will have an adverse effect on the model effect. Therefore, with a cumulative frequency of 100 as the critical point, 403 spectral feature variables were finally obtained. The classification accuracy trend chart with the increase of the number of variables is shown in the figure below. Figure 11 shown.

[0121] In order to verify the effectiveness of spectral variable selection, the near-infrared spectral feature variable adaptive selection algorithm was compared with full spectrum and spectral feature variable selection methods such as uninformative variable elimination (UVE) and particle swarm optimization (PSO). The tobacco aging degree discrimination model constructed by the present invention was used as the prediction model. The number of selected spectral feature variables and the classification accuracy of the test set were used as evaluation criteria. The results are shown in Table 3:

[0122] Table 3 Classification performance of each type of feature spectrum on the test set

[0123]

[0124] As can be seen from the table above, various feature variable selection methods reduce the number of spectral variables relative to the full spectrum, while also improving classification accuracy. The near-infrared spectral feature variable adaptive selection algorithm extracts the fewest spectral feature variables (spectral feature variables only account for about 25% of the full spectrum) and has the highest classification accuracy, indicating that the spectral feature variables extracted by this algorithm can effectively reduce redundant information and noise, making the model more robust and more generalizable.

[0125] Step (7), performance evaluation of the tobacco strip alcoholization degree discrimination model: use the tobacco strip alcoholization degree discrimination model to predict the test set and judge the model performance.

[0126] The accuracy rate is calculated for the entire model, which is the ratio of the number of correctly identified samples to the total number of observed samples. The higher the accuracy rate, the better the model classification performance. The calculation method is as follows:

[0127]

[0128] Among them, TP represents the number of samples whose true value is positive and the model determines it to be positive; TN represents the number of samples whose true value is negative and the model determines it to be negative; FP represents the number of samples whose true value is negative and the model determines it to be positive; FN represents the number of samples whose true value is positive and the model determines it to be negative.

[0129] Three commonly used near-infrared spectral classification algorithm models, PLS-DA, SVM-DA, and SIMCA, were selected to conduct a comparative analysis of classification performance with the tobacco aging degree discrimination model. The results are shown in Table 4:

[0130] Table 4 Performance comparison of different near-infrared spectral classification algorithms

[0131]

[0132] The alcoholization degree discrimination model of tobacco strips was used to predict the alcoholization degree of the test set samples and compared with the actual classification results. The comparison results are shown in Table 5:

[0133] Table 5 Comparison results

[0134] Sample number Model classification results Actual classification results Consistency determination CS0001 Unalcoholized expiration Unalcoholized expiration Consistency CS0002 Aging period Aging period Consistency CS0003 Aging period Aging period Consistency CS0004 perolation perolation Consistency … … … CS1545 Aging period Aging period Consistency

[0135] As can be seen in Table 5, the results of the present invention are consistent with those of the sensory evaluation panel. This demonstrates the accuracy and effectiveness of the present method, which can meet the needs of large-scale quality evaluation and determination in industrial production. The present method, through near-infrared spectroscopy and model-based determination, can obtain accurate and objective evaluation results, with sample testing and evaluation taking approximately one minute per sample.

[0136] After the model is trained and tested, the degree of aging of the tobacco strips is predicted, including: feature extraction of the acquired near-infrared spectral data of the tobacco strips to be tested to obtain a number of spectral characteristic variables to be tested; using a spectral variable selection algorithm to screen the number of spectral characteristic variables to be tested, and inputting the screened spectral characteristic variables to be tested into the tobacco strip aging degree discrimination model to obtain the degree of aging of the tobacco strips.

[0137] Compared to traditional sensory evaluation methods, the method developed in this paper for rapidly determining the degree of aging in stock tobacco strips comprehensively and objectively characterizes the quality of stock tobacco strips, resolving issues such as the subjectivity of sensory evaluation, unstable evaluation, and lack of credibility. The method is time-efficient, efficient, and low-cost.

[0138] Example 2

[0139] like Figure 12 As shown, this embodiment provides an adaptive search system for determining the alcoholization degree of tobacco chips in stock, including:

[0140] The model training module is configured to: obtain near-infrared spectral data and corresponding degrees of aging of inventory tobacco strips, perform feature extraction on the near-infrared spectral data of the inventory tobacco strips to obtain a number of spectral characteristic variables; use the several spectral characteristic variables as inputs to a tobacco strip aging degree discrimination model, use the corresponding degrees of aging as outputs, train the tobacco strip aging degree discrimination model, and calculate an objective function value based on the model output results and the corresponding degrees of aging; use a spectral variable selection algorithm to evaluate the quality of the spectral characteristic variables based on the objective function value, screen the spectral characteristic variables that meet the requirements, and re-input the spectral characteristic variables into the tobacco strip aging degree discrimination model for iterative training until the objective function value meets the set requirements;

[0141] The model prediction module is configured to: extract features from the acquired near-infrared spectral data of the tobacco strips to be tested to obtain a number of spectral characteristic variables to be tested; use a spectral variable selection algorithm to screen the number of spectral characteristic variables to be tested, input the screened spectral characteristic variables to be tested into the tobacco strip aging degree discrimination model to obtain the tobacco strip aging degree.

[0142] In some embodiments, the model training module is specifically configured to: set the search space of the cellular structure, select the cellular structure with the best performance and train it until convergence, obtain the search parameter set of the optimal structure, record the optimal weight, and assign it to all combined cellular structures in the one-dimensional convolutional neural network, use a cellular structure, hierarchically perform layer-by-layer search expansion in the superimposed and combined search space, and construct a model for distinguishing the degree of aging of tobacco leaves.

[0143] In some embodiments, the model training module is further configured to:

[0144] The search space of the cellular structure is formed by the value ranges of three hyperparameters: the convolution kernel size n1, the size of the downsampling kernel, and the number of feature maps N;

[0145] Normalize the value range of each hyperparameter to [0,1], randomly select a set of parameter configurations, and calculate the objective function EI(θ);

[0146] In order to deal with the mixed parameter space search range, a local neighborhood search range is initialized around the set of parameters, and the parameter values ​​θ′ of the S nearest neighbors are selected from the univariate Gaussian distribution as neighborhood samples;

[0147] For the S nearest neighbor samples, evaluate their objective function EI(θ′) at one time and select the neighbor sample parameter θ corresponding to the optimal objective function value inc , continue searching with it as the center;

[0148] Repeat the local search process and select the maximum EI(θ inc ) corresponds to the cell structure parameter θ inc , record the optimal cell structure weight v;

[0149] According to the obtained optimal hyperparameter combination, the cellular structure is expanded into a one-dimensional convolutional neural network.

[0150] In some embodiments, the model training module is further configured to:

[0151] Select the optimal cell structure weight ω obtained by the search, add it to the queue, and share this weight in the next layer;

[0152] Starting from the second layer, the model of each layer is expanded by adding all possible cell structures in the allowed search space, and all the models expanded in this layer are trained and evaluated;

[0153] The two network structures with the best prediction performance are selected and expanded into a complete one-dimensional convolutional neural network, which is added to the next layer queue; the network structure is retrained using the training data set and the weights of the previous layers are updated;

[0154] The cellular structure is added layer by layer from the second layer to the tenth layer, and the above steps are repeated until the preset search termination condition is reached to obtain a model for distinguishing the degree of alcoholization of tobacco leaves.

[0155] In some embodiments, the model training module / model prediction module is further configured to:

[0156] Initialize the harmony memory HM, initialize each component in the harmony vector, and record the objective function value of each component to form a harmony memory HMS×(n+1);

[0157] Compare the random number r1 between [0,1] and the probability HMCR of the harmony memory library to decide whether to select the harmony vector from the harmony memory library; compare the random number r2 between [0,1] and the pitch fine-tuning probability PAR, adjust the harmony vector according to the pitch fine-tuning bandwidth BW adaptive adjustment method, and generate a new harmony vector X new ;

[0158] According to the update of the Sigmoid function after fine-tuning the harmony variable, a random number r4 is generated between [0-1] and Sigmoid(x′ i ) for comparison, if r4<Sigmoid(x′ i ), x′ i The value is set to 1; if r4≥Sigmoid(x′ i ), x′ i The value of is set to 0 to represent each dimension of the harmony vector only with 0 or 1;

[0159] Evaluate the newly generated harmony vector X new , if f(X new ) is better than f(X worst ), then change the X in HM worst With the new harmony vector X new Replace; otherwise, X new Discard it directly, keep HM unchanged, and repeat the iteration until the number of iterations reaches Tmax.

[0160] In some embodiments, the model training module is further configured to calculate the mean classification accuracy as the objective function value based on the model output results and the corresponding degree of aging.

[0161] Example 3

[0162] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps of the method for adaptively searching for alcoholization degree of inventory tobacco sheets as described in the first embodiment above are implemented.

[0163] Example 4

[0164] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for determining the degree of alcoholization of tobacco leaves in inventory by adaptively searching for the program are implemented as described in the first embodiment above.

[0165] Example 5

[0166] This embodiment provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the method for determining the degree of alcoholization of tobacco leaves by adaptively searching inventory, as described in the first embodiment.

[0167] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer-usable program code.

[0168] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.

[0169] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0170] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0171] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0172] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for determining the degree of aging of tobacco leaves by adaptively searching for inventory, characterized in that: include: Obtaining near-infrared spectral data of the tobacco strips in stock and the corresponding degree of aging, performing feature extraction on the near-infrared spectral data of the tobacco strips in stock to obtain several spectral feature variables; Several spectral characteristic variables are used as inputs to the model for distinguishing the degree of alcoholization of tobacco strips, and the corresponding degree of alcoholization is used as output. The model for distinguishing the degree of alcoholization of tobacco strips is trained, and the objective function value is calculated based on the model output results and the corresponding degree of alcoholization. Using the spectral variable selection algorithm, the quality of the spectral characteristic variables is evaluated according to the objective function value, and the spectral characteristic variables that meet the requirements are screened and re-input into the tobacco aging degree discrimination model for iterative training until the objective function value meets the set requirements; Perform feature extraction on the acquired near-infrared spectrum data of the tobacco leaf to be detected, and obtain a number of spectrum feature variables to be detected; A spectral variable selection algorithm is used to screen several spectral characteristic variables to be detected, and the screened spectral characteristic variables to be detected are input into the discrimination model of the degree of alcoholization of tobacco leaves to obtain the degree of alcoholization of tobacco leaves.

2. The method for determining the degree of aging of tobacco chips by adaptively searching the inventory according to claim 1, characterized in that: The process of constructing the model for distinguishing the degree of aging of tobacco strips includes: setting a search space for a cellular structure, selecting the cellular structure with the best performance and training it until convergence, obtaining a set of search parameters for the optimal structure, recording the optimal weights, and assigning them to all combined cellular structures in a one-dimensional convolutional neural network. Using a cellular structure, layer-by-layer search expansion is performed in the superimposed and combined search space to construct a model for distinguishing the degree of aging of tobacco strips.

3. The method for determining the degree of aging of tobacco chips by adaptively searching the inventory according to claim 2, characterized in that: The method includes setting a search space for the cellular structure, selecting the cellular structure with the best performance and training it until convergence, obtaining a search parameter set for the optimal structure, recording the optimal weights, and assigning them to all combinations of cellular structures in the one-dimensional convolutional neural network. The search space of the cellular structure is formed by the value ranges of three hyperparameters: the convolution kernel size n1, the size of the downsampling kernel, and the number of feature maps N; Normalize the value range of each hyperparameter to [0,1], randomly select a set of parameter configurations, and calculate the objective function EI(θ); In order to deal with the mixed parameter space search range, a local neighborhood search range is initialized around the set of parameters, and the parameter values ​​θ′ of the S nearest neighbors are selected from the univariate Gaussian distribution as neighborhood samples; For the S nearest neighbor samples, evaluate their objective function EI(θ′) at one time and select the neighbor sample parameter θ corresponding to the optimal objective function value inc , continue searching with it as the center; Repeat the local search process and select the maximum EI(θ inc ) corresponds to the cell structure parameter θ inc , record the optimal cell structure weight ω; According to the obtained optimal hyperparameter combination, the cellular structure is expanded into a one-dimensional convolutional neural network.

4. The method for determining the degree of aging of tobacco chips by adaptively searching the inventory according to claim 3, characterized in that: The method uses a cellular structure to perform layer-by-layer search expansion in a superimposed and combined search space to construct a model for distinguishing the degree of alcoholization of tobacco leaves. The method includes: Select the optimal cell structure weight ω obtained by the search, add it to the queue, and share this weight in the next layer; Starting from the second layer, the model of each layer is expanded by adding all possible cell structures in the allowed search space, and all the models expanded in this layer are trained and evaluated; The two network structures with the best prediction performance are selected and expanded into a complete one-dimensional convolutional neural network, which is added to the next layer queue; the network structure is retrained using the training data set and the weights of the previous layers are updated; The cellular structure is added layer by layer from the second layer to the tenth layer, and the above steps are repeated until the preset search termination condition is reached to obtain a model for distinguishing the degree of alcoholization of tobacco leaves.

5. The method for determining the degree of aging of tobacco chips by adaptively searching the inventory according to claim 1, characterized in that: The spectral variable selection algorithm is used to evaluate the quality of spectral characteristic variables according to the objective function value, and to screen spectral characteristic variables that meet the requirements; the method includes: Initialize the harmony memory HM, initialize each component in the harmony vector, and record the objective function value of each component to form a harmony memory HMS×(n+1); Compare the random number r1 between [0,1] and the probability HMCR of the harmony memory library to decide whether to select the harmony vector from the harmony memory library; compare the random number r2 between [0,1] and the pitch fine-tuning probability PAR, adjust the harmony vector according to the pitch fine-tuning bandwidth BW adaptive adjustment method, and generate a new harmony vector X new ; According to the update of the Sigmoid function after fine-tuning the harmony variable, a random number r4 is generated between [0-1] and Sigmoid(x′ i ) for comparison, if r4<Sigmoid(x′ i ), x′ i The value is set to 1; if r4≥Sigmoid(x′ i ), x′ i The value of is set to 0 to represent each dimension of the harmony vector only with 0 or 1; Evaluate the newly generated harmony vector X new , if f(X new ) is better than f(X worst ), then change the X in HM worst With the new harmony vector X new Replace; otherwise, X new Discard it directly, keep HM unchanged, and repeat the iteration until the number of iterations reaches Tmax.

6. The method for determining the degree of aging of tobacco chips by adaptively searching the inventory according to claim 5, characterized in that: The objective function value is calculated according to the model output result and the corresponding degree of alcoholization; the method includes: calculating the classification accuracy mean according to the model output result and the corresponding degree of alcoholization as the objective function value.

7. An adaptive search system for determining the degree of aging of tobacco leaves in stock, characterized by: include: The model training module is configured to: obtain near-infrared spectral data of the stock tobacco strips and the corresponding degree of aging, perform feature extraction on the near-infrared spectral data of the stock tobacco strips, and obtain a plurality of spectral feature variables; Several spectral characteristic variables are used as inputs to the model for distinguishing the degree of alcoholization of tobacco strips, and the corresponding degree of alcoholization is used as output. The model for distinguishing the degree of alcoholization of tobacco strips is trained, and the objective function value is calculated based on the model output results and the corresponding degree of alcoholization. Using the spectral variable selection algorithm, the quality of the spectral characteristic variables is evaluated according to the objective function value, and the spectral characteristic variables that meet the requirements are screened and re-input into the tobacco aging degree discrimination model for iterative training until the objective function value meets the set requirements; The model prediction module is configured to: perform feature extraction on the acquired near-infrared spectral data of the tobacco leaf to be detected, and obtain a plurality of spectral feature variables to be detected; A spectral variable selection algorithm is used to screen several spectral characteristic variables to be detected, and the screened spectral characteristic variables to be detected are input into the discrimination model of the degree of alcoholization of tobacco leaves to obtain the degree of alcoholization of tobacco leaves.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the adaptive search method for determining the degree of alcoholization of stock tobacco slices as described in any one of claims 1 to 6 are implemented.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps in the adaptive search method for determining the degree of alcoholization of stock tobacco slices as described in any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that The computer program product includes a computer program, and when the computer program is executed by a processor, the steps in the adaptive search method for determining the alcoholization degree of stock tobacco slices as described in any one of claims 1 to 6 are implemented.

Images